Re: [PATCH] printk: Make console tracepoint safe in NMI() context

[Petr Mladek]

On Fri 2022-07-15 08:10:00, Paul E. McKenney wrote:
> On Fri, Jul 15, 2022 at 09:51:56AM -0400, Steven Rostedt wrote:
> > On Fri, 15 Jul 2022 14:39:52 +0200
> > Marco Elver <elver@xxxxxxxxxx> wrote:
> > 
> > > Couldn't this just use rcu_is_watching()?
> > > 
> > >   | * rcu_is_watching - see if RCU thinks that the current CPU is not idle
> > 
> > Maybe, but I was thinking that Petr had a way to hit the issue that we
> > worry about. But since the non _rcuide() call requires rcu watching,
> > prehaps that is better to use.

I actually saw the warning even with simple sysrq+l. I wonder why
I have missed it during testing. It was probably well hidden within
the other backtraces.

I was not aware that rcu_is_watching() and rcu_is_idle_cpu() did
basically the same. I used rcu_is_idle_cpu() because of the "idle"
in the name and the function description ;-)

> In case this helps...  ;-)
> 
> The rcu_is_watching() function is designed to be used from the current
> CPU, so it dispenses with memory ordering.  However, it explicitly
> disables preemption in order to avoid weird preemption patterns.
> 
> The formulation that Marco used is designed to be used from a remote
> CPU, and so it includes explicit memory ordering that is not needed
> in this case.  But it does not disable preemption.
> 
> So if preemption is enabled at that point in tracing, you really want
> to be using rcu_is_watching().

rcu_is_watching() is the right variant then. I am going to send v2.

Thanks a lot for the detailed explanation.

Best Regards,
Petr